Tromsø kommune
Tromsø municipality withdraws a school-closure proposal after ChatGPT invented its research sources
Staff in Tromsø, Norway, used ChatGPT from private accounts while drafting the research chapter of a 2025 proposal to restructure schools and kindergartens. The sources it cited did not exist. The municipality apologised, withdrew the proposal, and a PwC review it commissioned found it was not mature enough to use AI.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
only seven of 18
Entries in the proposal's source list that could be traced, in the review by the local newspaper iTromsø (reported by digi.no and NTB)1
7
Sources in chapter 11.4 of the proposal that PwC lists and confirms were invented by the AI2
half a year
Length of time the municipality's chief executive halted the consultation on the proposal1
What it took
The problem
In February 2025 Tromsø municipality in northern Norway sent out for public consultation a proposal for a new kindergarten and school structure, with eight schools proposed for closure. Its knowledge base included a research chapter. Shortly after publication, there were accusations that some of its sources did not exist.
What they built
This was not a planned AI project. The project group had been encouraged to use AI to ease the writing, and staff mainly used ChatGPT from private accounts. According to PwC, ChatGPT was used to summarise a prepared text about school sizes and was then asked for sources. It supplied references to research and publications that do not exist, an effect called hallucination.
How they did it
The AI-written text and references went into the consultation document without being caught by ordinary quality checks. Professors named in the list said they had never written the works credited to them. The municipality's chief executive said people wrote the knowledge base and that AI was used as a tool. The municipality admitted the sources did not exist, apologised, and stopped the consultation for half a year. It then commissioned PwC to review the project and its use of AI.
Results
The failure had real costs. The local newspaper iTromsø found only seven of 18 listed sources traceable. The municipality withdrew the case and hired an external review. PwC presented its report on 2 July 2025. It confirmed that seven sources in the research chapter were invented by the AI. It found no evidence the errors changed the proposal, but said they could damage trust.
What went wrong
PwC found the municipality had no AI strategy, unclear responsibility for AI, guidelines that named no approved systems, and no training for staff. The AI-generated text was not quality-checked, and it was not labelled as AI-generated. PwC concluded the municipality was not mature enough to adopt AI.
What to copy
Before staff use a general chatbot on public documents, set rules for which tools are allowed, require a person to verify every source, and label AI-generated text.
Lessons for your project
Never let staff ask a chatbot for sources and paste them in. Treat every reference an AI gives as unverified until a person has found the original.
Write down which AI tools staff may use before they use them. PwC found Tromsø had no AI strategy, unclear responsibility, and rules that named no approved systems.
Train staff and label AI-written text. PwC found staff were left to teach themselves, and the generated text was neither checked nor marked as AI-generated.
Build a check into high-stakes papers. A public consultation document needs a named reviewer who confirms every cited source exists.
Caveats
- This is a failure case. The municipality did not publish a result from the AI. The sources are the municipality's own page, a review by PwC that it commissioned, and news reports.
- The count of unverifiable sources differs by source. The newspaper review found seven of 18 entries traceable. PwC lists seven invented sources in one chapter. Neither source's count is combined or recomputed here.
- The digi.no article only quotes a professor and a researcher suspecting ChatGPT. PwC's report confirms it, stating it was mainly used from private accounts, and the municipality had sometimes paid for individual accounts with external providers after a manager approved them. This is staff use that the project group was encouraged to try.
- PwC found no evidence that the fabricated sources affected the proposal's conclusion, because the direction was already set by a 2021 council decision, though the errors could still reduce trust in the municipality.
- PwC says its descriptions and assessments are its own, and that the AI policy was set up midway through the project, so it is unclear whether it applied when ChatGPT was used.
- The PwC report returned an access block on pwc.no and was read from a web archive copy. The deployed date is the month the proposal went out for consultation. The municipality's page prints "Publisert 27. jun 2025" (last updated 2 July 2025), and that is the date used for it. The 1.2 million kroner figure is a media report, not PwC or the municipality.
Sources
- Skandalen i Tromsø: Kommunen brukte kilder som KI hadde diktet opp
Digi.no (NTB) · 2025-03-28 · News · digi.no
- Evaluering av bruk av kunstig intelligens i Tromsø kommune
PwC (commissioned by Tromsø kommune) · 2025-06-30 · Filing · pwc.no
- KI-konklusjon klar: Tromsø kommune var ikke moden for å ta i bruk KI
NRK Troms og Finnmark · 2025-07-02 · News · nrk.no
- Her er rapporten om KI-bruk og barnehage- og skolestruktur
Tromsø kommune · 2025-06-27 · Press release · tromso.kommune.no
Summarized in our own words from the sources above. Spotted an error? See corrections.